Research on cabbage transplanting status detection and operation quality evaluation in complex environments based on improved YOLOv10-TQ and DeepSort.

The quality of crop transplanting is a critical factor influencing both plant survival and final yield. However, in mechanized transplanting operations, manual inspection suffers from low efficiency, while traditional algorithms struggle with poor environmental adaptability and insufficient detection accuracy. To address these limitations, this study proposes a detection-and-tracking-based method for recognizing and counting cabbage transplanting states in open-field scenarios, enabling accurate identification and robust tracking of seedling transplanting conditions. An improved YOLOv10-TQ detection network is developed by integrating a triplet attention mechanism and a combined QFocal Loss-cross-entropy loss function to enhance the detection accuracy and stability for three transplanting states of cabbage: normal, soil-buried seedlings, and bare-root seedlings. In addition, a lightweight MobileViT feature extraction network is incorporated into the DeepSort algorithm to improve fine-grained target representation, and, combined with a line-crossing counting strategy, this approach enables identity de-duplication and robust counting performance. Experimental results demonstrate that the proposed method achieves a mean average precision (mAP) of 86.3% and an average counting accuracy of 97.8% on a self-constructed cabbage transplanting dataset. Based on this study, a visualization system for monitoring cabbage transplanting status was designed to enhance precision in agricultural practices. Compared with traditional detection and counting methods, the proposed approach exhibits significant advantages in detection accuracy, tracking stability, and counting precision, providing a promising technical foundation for intelligent quality evaluation of cabbage transplanting operations and data-driven decision-making in agricultural machinery systems.